Characterization of physically based hydrologic model behaviour with temporal sensitivity analysis for flash floods in Mediterranean catchments

This paper presents a detailed analysis of 10 flash flood events in the Mediterranean region using the distributed hydrological model MARINE. Characterizing catchment's response during flash flood events may provide a new and valuable insight into the processes involved for extreme flood response and their dependency on catchment properties and flood severity. The main objective of this study is to analyze hydrologic model sensitivity in the case of flash floods with a new approach in hydrology, allowing model outputs variance decomposition for temporal patterns of parameter sensitivity analysis. Such approaches enable ranking of uncertainty sources for non-linear and non-monotonic mappings with a low computational cost. This study uses hydrologic model and sensitivity analysis as learning tools to derive temporal sensitivity analysis with a variance based method in the case of 10 flash floods that occurred in the French Pyrenees and Cévennes foothills. This constitutes a huge dataset given the scarcity of data about flash flood events. With Nash performances above 0.73 on average for this extended set of validation events, the five sensitive parameters of MARINE distributed physically based model are analyzed. This contribution shows that soil depth explains more than 80% of model output variance when most hydrographs are peaking. Moreover the lateral subsurface transfer is responsible for 80% of model variance for some catchment-flood events' hydrographs during slow declining limbs. The unexplained variance of model output representing interactions between parameters reveals to be very low during modeled flood peaks and informs that model parsimonious parameterization is appropriate to tackle the problem of flash floods. Interactions observed after model initialization or rainfall intensity peaks incite to improve water partition representation between flow components and initialization itself. This paper gives a practical framework for application of this method to other models, landscapes and climatic conditions, potentially helping to improve processes understanding and representation.